Google DeepMind has launched Gemma 3, a new collection of state-of-the-art open models designed to bring advanced AI capabilities to a wider range of devices. Built on the same research and technology powering the Gemini 2.0 models, Gemma 3 is engineered for speed and efficiency, enabling developers to run powerful AI applications directly on single GPUs and TPUs.
This release marks a significant advancement in accessible AI, building on the success of the Gemma family, which celebrated its first anniversary with over 100 million downloads and a thriving community known as the “Gemmaverse.”
Gemma 3: Advanced Capabilities for Diverse Applications
Gemma 3 comes in various sizes (1B, 4B, 12B, and 27B), allowing developers to select the optimal model for their specific hardware and performance needs. Key features include:
- Superior Performance: Gemma 3 27B has demonstrated exceptional performance, outperforming models like Llama-405B and DeepSeek-V3 in preliminary human preference evaluations, while requiring only a single GPU.
- Multilingual Support: With out-of-the-box support for over 35 languages and pretrained support for over 140, Gemma 3 enables the development of global applications.
- Advanced Reasoning: The models offer advanced text and visual reasoning capabilities, allowing for the analysis of images, text, and short videos.
- Expanded Context Window: A 128k-token context window enables the processing of vast amounts of information.
- Function Calling: Support for function calling and structured output facilitates the automation of tasks and the creation of agentic experiences.
- Quantized Models: Official quantized versions enhance performance by reducing model size and computational requirements.
Commitment to Responsible AI Development
Google DeepMind emphasizes its commitment to responsible AI development, implementing rigorous safety protocols for Gemma 3. This includes extensive data governance, alignment with safety policies through fine-tuning, and robust benchmark evaluations.
ShieldGemma 2: Enhanced Image Safety
Alongside Gemma 3, Google DeepMind introduced ShieldGemma 2, a 4B image safety checker built on the Gemma 3 foundation. ShieldGemma 2 provides safety labels across three categories: dangerous content, sexually explicit, and violence, offering developers a customizable solution for image safety.
Seamless Integration and Deployment
Gemma 3 and ShieldGemma 2 are designed for seamless integration with popular tools and platforms, including Hugging Face Transformers, Ollama, JAX, Keras, PyTorch, Google AI Edge, UnSloth, vLLM, and Gemma.cpp. Developers can experiment with Gemma 3 in Google AI Studio, or download models from Kaggle and Hugging Face.
Deployment options include Vertex AI, Cloud Run, the Google GenAI API, local environments, and other platforms. NVIDIA has optimized Gemma 3 for its GPUs, and the models are also compatible with Google Cloud TPUs and AMD GPUs.
Fostering Innovation with the Gemmaverse
The Gemmaverse, a community-driven ecosystem of models and tools, continues to expand, showcasing the potential of Gemma in diverse applications. Google DeepMind is also launching the Gemma 3 Academic Program, offering Google Cloud credits to accelerate academic research.
Getting Started with Gemma 3
Developers can begin exploring Gemma 3 through Google AI Studio, download models from various platforms, and customize them to meet their specific needs. Google DeepMind encourages developers to leverage the power of Gemma 3 to create innovative AI applications.
